llama3.1-cpo_j-full-0912

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the princeton-nlp/llama3-ultrafeedback dataset. It achieves the following results on the evaluation set:

• Loss: 1.4395
• Rewards/chosen: -16.1609
• Rewards/rejected: -16.9344
• Rewards/accuracies: 0.6326
• Rewards/margins: 0.7735
• Logps/rejected: -169.3439
• Logps/chosen: -161.6093
• Logits/rejected: -0.3578
• Logits/chosen: -0.3883
• Nll Loss: 0.2841

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Nll Loss
1.7848	0.2311	100	1.6452	-15.3752	-15.7662	0.5804	0.3910	-157.6625	-153.7521	-0.3516	-0.3794	0.2719
1.5276	0.4623	200	1.5229	-15.8100	-16.4430	0.6043	0.6331	-164.4303	-158.0997	-0.3983	-0.4237	0.2748
1.4811	0.6934	300	1.4640	-16.0706	-16.8001	0.6130	0.7296	-168.0013	-160.7057	-0.4069	-0.4339	0.2804
1.4642	0.9246	400	1.4429	-16.1577	-16.9120	0.6304	0.7544	-169.1204	-161.5765	-0.3509	-0.3812	0.2845

Framework versions

• Transformers 4.44.2
• Pytorch 2.3.1
• Datasets 2.21.0
• Tokenizers 0.19.1